Reading. 2. Fourier analysis and sampling theory. Required: Watt, Section 14.1 Recommended:

Transcription

1 Reading Required: Watt, Section 14.1 Recommended: 2. Fourier analysis and sampling theory Ron Bracewell, The Fourier Transform and Its Applications, McGraw-Hill. Don P. Mitchell and Arun N. Netravali, Reconstruction Filters in Computer Computer Graphics, Computer Graphics, (Proceedings of SIGGRAPH 88). 22 (4), pp ,

2 What is an image? Images as functions We can think of an image as a function, f, from R 2 to R: f( x, y ) gives the intensity of a channel at position ( x, y ) Realistically, we expect the image only to be defined over a rectangle, with a finite range: f: [a,b]x[c,d] [0,1] A color image is just three functions pasted together. We can write this as a vector-valued function: rxy (, ) f( x, y) = g( x, y) bxy (, ) We ll focus in grayscale (scalar-valued) images for now. 3 4

3 Digital images In computer graphics, we usually create or operate on digital (discrete) images: Sample the space on a regular grid Quantize each sample (round to nearest integer) If our samples are apart, we can write this as: f[i,j] = Quantize{ f(i, j ) } Motivation: filtering and resizing What if we now want to: smooth an image? sharpen an image? enlarge an image? shrink an image? Before we try these operations, it s helpful to think about images in a more mathematical way 5 6

4 Fourier transforms We can represent functions as a weighted sum of sines and cosines. We can think of a function in two complementary ways: Spatially in the spatial domain Spectrally in the frequency domain The Fourier transform and its inverse convert between these two domains: Fourier transforms (cont d) Spatial domain i2π sx Fs () = f( xe ) dx 2 f( x) Fse ( ) i π sx ds = Where do the sines and cosines come in? Frequency domain Spatial domain i2π sx Fs () f( xe ) dx = 2 f( x) Fse ( ) i π sx ds = Frequency domain f(x) is usually a real signal, but F(s) is generally complex: Fs () = As () + ibs () = Fs () e i2 πθ ( s) If f(x) is symmetric, i.e., f(x) = f(-x)), then F(s) = A(s). Why? 7 8

5 1D Fourier examples 2D Fourier transform Spatial domain i2 π ( s x+ s y) x y Fs (, s) f( x, ye ) dxdy x y = i2 π ( s x+ s y) x y x y x y f( x, y) Fs (, s ) e dsds = Frequency domain Spatial domain Frequency domain f( x, y) Fs (, s) x y 9 10

6 2D Fourier examples Convolution One of the most common methods for filtering a function is called convolution. In 1D, convolution is defined as: Spatial domain f( x, y) Frequency domain Fs (, s) x y gx ( ) = f( x) hx ( ) = f( x') h( x x') dx' = f( x') h% ( x' x) dx' where hx %( ) h( x)

7 Convolution properties Convolution in 2D Convolution exhibits a number of basic, but important properties. Commutativity: Associativity: ax ( ) bx ( ) = bx ( ) ax ( ) [ ax ( ) bx ( )] cx ( ) = ax ( ) [ bx ( ) cx ( )] In two dimensions, convolution becomes: gxy (, ) = f( xy, ) hxy (, ) = f ( x', y') h( x x', y y') dx' dy' = f ( x', y') h% ( x' x, y' y) dx' dy' where hxy %(, ) = h( x, y). Linearity: ax ( ) [ k bx ( )] = k [ ax ( ) bx ( )] ax ( ) ( bx ( ) + cx ( )) = ax ( ) bx ( ) + ax ( ) cx ( ) * = f(x,y) h(x,y) g(x,y) 13 14

8 Convolution theorems 1D convolution theorem example Convolution theorem: Convolution in the spatial domain is equivalent to multiplication in the frequency domain. f h F H Symmetric theorem: Convolution in the frequency domain is equivalent to multiplication in the spatial domain. f h F H 15 16

9 2D convolution theorem example The delta function f(x,y) F(s x,s y ) The Dirac delta function, δ(x), is a handy tool for sampling theory. It has zero width, infinite height, and unit area. It is usually drawn as: * h(x,y) H(s x,s y ) g(x,y) G(s x,s y ) 17 18

10 Sifting and shifting For sampling, the delta function has two important properties. Sifting: f( x) δ ( x a) = f( a) δ ( x a) The shah/comb function A string of delta functions is the key to sampling. The resulting function is called the shah or comb function: which looks like: III( x) = δ ( x nt) n= Shifting: f( x) δ ( x a) = f( x a) Amazingly, the Fourier transform of the shah function takes the same form: where s o = 1/T. III( s) = δ ( s ns ) n= o 19 20

11 Sampling Sampling and reconstruction Now, we can talk about sampling. The Fourier spectrum gets replicated by spatial sampling! How do we recover the signal? 21 22

12 Sampling and reconstruction in 2D Sampling theorem This result is known as the Sampling Theorem and is due to Claude Shannon who first discovered it in 1949: A signal can be reconstructed from its samples without loss of information, if the original signal has no energy in frequencies at or above ½ the sampling frequency. For a given bandlimited function, the minimum rate at which it must be sampled is the Nyquist frequency

13 Reconstruction filters The sinc filter, while ideal, has two drawbacks: It has large support (slow to compute) It introduces ringing in practice We can choose from many other filters Cubic filters Mitchell and Netravali (1988) experimented with cubic filters, reducing them all to the following form: 3 2 B C x + + B+ C x + B x < ( ) ( ) (6 2 ) rx ( ) = (( B 6 Cx ) + (6B+ 30 Cx ) + ( 12B 48 Cx ) + (8B+ 24 C) 1 x< otherwise The choice of B or C trades off between being too blurry or having too much ringing. B=C=1/3 was their visually best choice. The resulting reconstruction filter is often called the Mitchell filter

14 Reconstruction filters in 2D Aliasing We can also perform reconstruction in 2D Sampling rate is too low 27 28

15 Aliasing What if we go below the Nyquist frequency? Anti-aliasing Anti-aliasing is the process of removing the frequencies before they alias

16 Anti-aliasing by analytic prefiltering We can fill the magic box with analytic prefiltering of the signal: Filtered downsampling Alternatively, we can sample the image at a higher rate, and then filter that signal: Why may this not generally be possible? We can now sample the signal at a lower rate. The whole process is called filtered downsampling or supersampling and averaging down

17 Practical upsampling Practical upsampling When resampling a function (e.g., when resizing an image), you do not need to reconstruct the complete continuous function. For zooming in on a function, you need only use a reconstruction filter and evaluate as needed for each new sample. Here s an example using a cubic filter: This can also be viewed as: 1. putting the reconstruction filter at the desired location 2. evaluating at the original sample positions 3. taking products with the sample values themselves 4. summing it up Important: filter should always be normalized! 33 34

18 Practical downsampling Downsampling is similar, but filter has larger support and smaller amplitude. Operationally: 1. Choose filter in downsampled space. 2. Compute the downsampling rate, d, ratio of new sampling rate to old sampling rate 3. Stretch the filter by 1/d and scale it down by d 4. Follow upsampling procedure (previous slides) to compute new values 2D resampling We ve been looking at separable filters: r ( x, y) = r ( x) r ( y) 2D 1D 1D How might you use this fact for efficient resampling in 2D? 35 36

19 Next class: Image Processing Reading: Jain, Kasturi, Schunck, Machine Vision. McGraw-Hill, Sections , 4.5(intro), 4.5.5, 4.5.6, (from course reader) Topics: - Implementing discrete convolution - Blurring and noise reduction - Sharpening - Edge detection 37